A FRAMEWORK FOR EVALUATING MACHINE AND DEEP LEARNING ALGORITHMS FOR LAND USE AND LAND COVER CLASSIFICATION FOR MEDIUM RESOLUTION IMAGERIES

Abstract

Land Use and Land Cover (LULC) information over large areas are essential for land resource management, environmental sustainability, impact of human activities and climate change etc. Several studies have evaluated and proposed the efficacy of Machine learning (ML) and Deep Learning (DL) models to classify LULC information but with no consensus on the choice of the best model for complex (spatially and spectrally similar land covers) LULC classification problems. This study used medium resolution remote sensing Sentinel-2 (S2) imagery to evaluate existing ML and DL algorithms and traditional Maximum Likelihood Classification (MLC) method for LULC classification. For the classification, two data sets, four bands and ten bands’ composites, and training and testing data were created accordingly. Four band composite included 10m spatial resolution bands, and ten bands composite, six other band including two SWIR, and four red edge bands were used respectively to evaluate their effect on classification accuracy, maximum class separability and computational power. For DL, the training and testing data in the form of patches and pixels were fed to the Convolutional Neural Network (CNN), Dl and MLC algorithms. The results show that the CNN model with overall classification accuracy of 97.5%, outperformed ML models (95.8%) and MLC classification (89 %). Among ML models, Random Forest achieved better accuracy (94%). The MLC achieved the lowest overall classification accuracy of 86%. All models produced better classification results and class separability on 4-bands composite than 10- bands, with not much computational time and resource. The results analysis show that CNN model can be more suitable for the classification of complex structured areas with spatially and spectrally confusing land covers.